Weight value fixed-point quantification method for lightweight convolutional neural network

Abstract

The invention discloses a weight value fixed-point quantization method for a lightweight convolutional neural network. The method comprises the steps of dividing weights, extracting a quantization strategy from a specific set, training, completing multi-measurement iteration, obtaining a quantization factor and completing quantization. Compared with the prior art, the method has the advantages that the retraining speed is higher, the accuracy loss is lower on the lightweight convolutional neural network, and the hardware resource consumption is lower than that of convolutional 8-bit integer multiplication.

Description

technical field [0001] The invention belongs to the field of lightweight convolutional neural networks, and in particular relates to a weighted numerical fixed-point quantization method for lightweight convolutional neural networks. Background technique [0002] With the maturity of deep neural network technology in the field of computer vision, the classification ability of deep convolutional network for specific objects has approached the level of human eye classification. Convolutional neural network (CNN) is becoming more and more mature in the field of computer vision. In the competition, the convolutional neural network algorithm is far more effective than the traditional algorithm. Therefore, the convolutional neural network algorithm has become a commonly used image processing algorithm. It extracts image features through a large number of floating-point multiplication and addition operations. In many hardware devices, the floating-point arithmetic unit is relatively...

AI Technical Summary

Problems solved by technology

The most simple and direct method is to quantize the weights and inputs into low-precision integers, such as 8-bit or 16-bit integers, and use low-bit-width integer operations to save hardware resources, but this will lead to a loss in the accuracy of the convolutional neural network, while The weight of the lightweight neural network is extremely small, and the parameter redundancy is low, resulting in a more obvious loss of accuracy. Therefore, how to effectively avoid the serious loss of accuracy is the key to the quantization of convolutional neural networks.

Images